How to Calculate AI Share of Voice with a Declared Denominator

Percentages travel further than their context does. 'Our AI share of voice hit 40%' will be repeated in three meetings by people who never saw the denominator — which is exactly why the denominator has to be stapled to the number every time it's written down.

The formula, precisely

SoV(window) = answers_mentioning_brand / total_sampled_answers

where total_sampled_answers =
  |prompt_set(version)| x |engines| x runs_per_prompt

Report as: 'X% (prompt set v2, 25 prompts, 2 engines, 3 runs, Sept 9-10)'

Mentions and citations get separate SoV numbers if you report both — never blended. And competitor SoV uses the same denominator on the same stored answers, or it isn't comparable.

A worked example — fictional, for shape only

Lumina Desk (fictional, illustrative numbers): prompt set v1, 25 prompts, 2 engines, 3 runs → 150 sampled answers per window. Window B logged 6 answers mentioning the brand: 6/150 = 4% mention-SoV, reported as '4% (v1, 25×2×3, Sept 9–10)'. A competitor named in 30 of the same answers sits at 20% on the identical denominator — a real gap, honestly measured, and a much more useful sentence than either number alone.

A failure worth checking

Denominator inflation, the reporting-season classic: only browsing-mode runs 'count' this quarter because they mention you more, or failed runs are silently dropped from the total. Every exclusion shrinks the denominator and flatters the percentage. The rule: exclusions are declared next to the number too, or the run stays in the denominator as a non-mention.

Limitations

Sampled answers are a sample: engines vary between identical runs, so small SoV movements inside a window are noise, not trend. Treat week-over-week shifts of a few points on a 150-answer sample as within variance until repeated across windows.